A comparison of probabilistic classifiers for sleep stage classification.

Area of Science:

• Computational neuroscience
• Biomedical engineering
• Sleep medicine

Background:

• Accurate sleep stage classification is crucial for diagnosing sleep disorders.
• Traditional methods struggle with complex or irregular sleep architectures.
• Obstructive sleep apnea (OSA) presents unique challenges for sleep staging.

Purpose of the Study:

• To compare Conditional Random Fields (CRF), Hidden Markov Models (HMMs), and Linear Discriminants (LDs) for cardiorespiratory sleep stage classification.
• To evaluate the impact of incorporating temporal information into classification algorithms.
• To assess the feasibility of sleep staging in OSA patients using cardiorespiratory data.

Main Methods:

• 10-fold cross-validation was employed on polysomnography (PSG) recordings from 231 participants (180 healthy, 51 with OSA).
• Classifiers were evaluated on five-, four-, and three-class sleep staging tasks.
• Conditional Random Fields with time information (CRFt) were compared against other models.

Main Results:

• CRFt significantly outperformed HMMs and LDs across all classification tasks and participant groups.
• CRFt achieved median accuracies ranging from 62.8% (5-class) to 77.6% (3-class) for all participants.
• Improved classification performance was observed for both healthy and OSA participants when using CRFt.

Conclusions:

• CRFt demonstrates superior ability in classifying complex and irregular sleep architectures.
• Sleep stage classification in OSA patients using cardiorespiratory features and CRFt is feasible with reasonable accuracy.
• CRFt offers a promising approach for advancing sleep staging in both clinical and research settings.